Bootstrap clustering for graph partitioning

Philippe Gambette (1) , Alain Guénoche (2)

1 LIGM - Laboratoire d'Informatique Gaspard-Monge
2 MMG - Marseille medical genetics - Centre de génétique médicale de Marseille

Given a simple undirected weighted or unweighted graph, we try to cluster the vertex set into communities and also to quantify the robustness of these clusters. For that task, we propose a new method, called bootstrap clustering which consists in (i) defining a new clustering algorithm for graphs, (ii) building a set of graphs similar to the initial one, (iii) applying the clustering method to each of them, making a profile (set) of partitions, (iv) computing a consensus partition for this profile, which is the final graph partitioning. This allows to evaluate the robustness of a cluster as the average percentage of partitions in the profile joining its element pairs ; this notion can be extended to partitions. Doing so, the initial and consensus partitions can be compared. A simulation protocol, based on random graphs structured in communities is designed to evaluate the efficiency of the Bootstrap Clustering approach.

Keywords: graph partitioning; clustering; modularity; consensus of partitions; bootstrap

Domains: Discrete Mathematics [cs.DM]

Philippe Gambette, Alain Guénoche. Bootstrap clustering for graph partitioning. RAIRO - Operations Research, 2011, 45 (4), pp.339-352. ⟨10.1051/ro/2012001⟩. ⟨hal-00676989⟩